Human approval on agent actions: what fraction of your approvals are rubber stamps?
retained text
Genuine question for anyone running a human-in-the-loop gate on agent actions. The standard advice is: let the agent draft, have a human approve anything risky. It sounds safe. But I'm curious how it holds up in practice. When you look at your approval queue honestly: what fraction of approvals are real catches versus autopilot clicks? Does the catch rate change over time as the approver learns to trust the agent, and is that trust earned or just fatigue? What does your queue show the approver: just the action, or the evidence behind it? And has anyone tried to measure this, even roughly? Asking because a gate nobody actually reads is just latency with extra steps, but I rarely see anyone talk about the approval side of the loop.
document fetched 2026-10-21 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xayuo8/human_approval_on_agent_actions_what_fraction_of